Changeset 75c7252 for libcfa

Timestamp:

Sep 23, 2021, 2:43:44 PM (3 years ago)

Author:

Thierry Delisle <tdelisle@…>

Branches:

ADT, ast-experimental, enum, forall-pointer-decay, master, pthread-emulation, qualifiedEnum

Children:

716b62c, 814f87d

Parents:

24e321c

Message:

CPU workstealing now has correct remote unpark.

Location:

libcfa/src/concurrency

Files:

• clib/cfathread.cfa (modified) (2 diffs)
• ready_queue.cfa (modified) (6 diffs)

libcfa/src/concurrency/clib/cfathread.cfa

@@ -129 +129 @@
                                 for(i; ret) {
                                         thread$ * thrd = (thread$ *)events[i].data.u64;
-                                        unpark( thrd );
+                                        unpark( thrd, UNPARK_REMOTE );
                                 }
                         }
@@ -158 +158 @@
                 }
 
-                poller_cnt = 24;
+                poller_cnt = 2;
                 poller_fds = alloc(poller_cnt);
                 pollers    = alloc(poller_cnt);

libcfa/src/concurrency/ready_queue.cfa

@@ -100 +100 @@
         #define __kernel_rseq_unregister rseq_unregister_current_thread
 #elif defined(CFA_HAVE_LINUX_RSEQ_H)
-        void __kernel_raw_rseq_register  (void);
-        void __kernel_raw_rseq_unregister(void);
+        static void __kernel_raw_rseq_register  (void);
+        static void __kernel_raw_rseq_unregister(void);
 
         #define __kernel_rseq_register __kernel_raw_rseq_register
@@ -294 +294 @@
 
                 processor * const proc = kernelTLS().this_processor;
-                const bool external = !push_local || (!proc) || (cltr != proc->cltr);
-
+                const bool external = (!proc) || (cltr != proc->cltr);
+
+                // Figure out the current cpu and make sure it is valid
                 const int cpu = __kernel_getcpu();
                 /* paranoid */ verify(cpu >= 0);
@@ -301 +302 @@
                 /* paranoid */ verify(cpu * READYQ_SHARD_FACTOR < lanes.count);
 
-                const cpu_map_entry_t & map = cpu_info.llc_map[cpu];
+                // Figure out where thread was last time and make sure it's
+                /* paranoid */ verify(thrd->preferred >= 0);
+                /* paranoid */ verify(thrd->preferred < cpu_info.hthrd_count);
+                /* paranoid */ verify(thrd->preferred * READYQ_SHARD_FACTOR < lanes.count);
+                const int prf = thrd->preferred * READYQ_SHARD_FACTOR;
+
+                const cpu_map_entry_t & map;
+                choose(hint) {
+                        case UNPARK_LOCAL : &map = &cpu_info.llc_map[cpu];
+                        case UNPARK_REMOTE: &map = &cpu_info.llc_map[prf];
+                }
                 /* paranoid */ verify(map.start * READYQ_SHARD_FACTOR < lanes.count);
                 /* paranoid */ verify(map.self * READYQ_SHARD_FACTOR < lanes.count);
@@ -312 +323 @@
                         if(unlikely(external)) { r = __tls_rand(); }
                         else { r = proc->rdq.its++; }
-                        i = start + (r % READYQ_SHARD_FACTOR);
+                        choose(hint) {
+                                case UNPARK_LOCAL : i = start + (r % READYQ_SHARD_FACTOR);
+                                case UNPARK_REMOTE: i = prf   + (r % READYQ_SHARD_FACTOR);
+                        }
                         // If we can't lock it retry
                 } while( !__atomic_try_acquire( &lanes.data[i].lock ) );
@@ -965 +979 @@
         extern void __enable_interrupts_hard();
 
-        void __kernel_raw_rseq_register  (void) {
+        static void __kernel_raw_rseq_register  (void) {
                 /* paranoid */ verify( __cfaabi_rseq.cpu_id == RSEQ_CPU_ID_UNINITIALIZED );
 
@@ -983 +997 @@
         }
 
-        void __kernel_raw_rseq_unregister(void) {
+        static void __kernel_raw_rseq_unregister(void) {
                 /* paranoid */ verify( __cfaabi_rseq.cpu_id >= 0 );